On Thu, 21 Mar 2019 09:33:17 +0100
Peter Zijlstra <pet...@infradead.org> wrote:
> On Wed, Mar 20, 2019 at 10:15:34PM -0400, Steven Rostedt wrote:
>
> > And it would crash similarly each time I tried it, but always at a
> > different place. After spending the day on this, I finally figured it
> > out. The bug is happening in entry_64.S right after error_entry.
> > There's two TRACE_IRQS_OFF in that code path, which if I comment out,
> > the bug goes away. Then it dawned on me that the crash always happens
> > when systemd does a normal page fault. We had this bug before, and it
> > was with the exception trace points.
>
> 0ac09f9f8cd1 ("x86, trace: Fix CR2 corruption when tracing page faults")
> d4078e232267 ("x86, trace: Further robustify CR2 handling vs tracing")
Probably these two, as I remember more about the discussions around
them, and not the actual commits. Although, I did take a look at the
do_page_fault() code that was added because of them. I just didn't do a
git blame to see what added it.
>
> Or were you talking about:
>
> 70fb74a5420f ("x86: Save cr2 in NMI in case NMIs take a page fault (for
> i386)")
>
> > The issue is that a tracepoint can fault (reading vmalloc or whatever).
> > And doing a userspace stack trace most definitely will fault. But if we
> > are coming from a legitimate page fault, the address of that fault (in
> > the CR2 register) will be lost if we fault before we get to the page
> > fault handler. That's exactly what is happening.
>
> Shees, that could've been written much clearer. So you're saying:
I wrote this just before going to bed. It was the best I could come up
with at the time.
>
> idtentry page_fault do_page_fault has_error_code=1
> call error_entry
> TRACE_IRQS_OFF
> call trace_hardirqs_off*
> <tracer stuff>
> <fault> # modifies CR2
> call do_page_fault
> address = read_cr2(); /* whoopsie */
>
> Right?
Yes.
>
> > To solve this, a TRACE_IRQS_OFF_CR2 (and ON for consistency) was added
> > that saves the CR2 register. A new trace_hardirqs_off_thunk_cr2 is
> > created that stores the cr2 register, calls the
> > trace_hardirqs_off_caller, then on return restores the cr2 register if
> > it changed, before returning.
>
> Yuck.. also, not consistent with the actual patch. The thunk doesn't
> save/restore CR2.
Well, the thunk calls the caller_cr2 that does, which is just a helper
function for the thunk.
>
> I really hate making this special TRACE_IRQS_OFF_CR2 thing, it feels far
> too fragile. I'd _much_ rather push the #PF CR2 read much earlier.
>
> Also, argh I fscking hate context tracking. That makes all this so much
> more complicated. It if weren't for CALL_enter_from_user_mode, we could
> pull that TRACE_IRQS_OFF out of error_entry.
Yeah, and I didn't even test this with context tracking enabled yet.
-- Steve
>
> Damn... Andy, any bright ideas?
